In [1]:
import sys
print sys.version
from joblib import Parallel, delayed
import multiprocessing
nCores = multiprocessing.cpu_count() - 2 # Allow other apps to run
print 'nCores: %d' % (nCores)
from datetime import datetime, time
print 'now: %s' % str(datetime.now())
In [2]:
import matplotlib.pyplot as plt
%matplotlib inline
from IPython.display import display, Image
import rpy2.robjects as robjects
from rpy2.robjects.packages import importr
from rpy2.robjects.lib import grid
from rpy2.robjects.lib import ggplot2
import rpy2.robjects.pandas2ri
import numpy as np
np.set_printoptions(precision=4, suppress=True)
import os
import pandas as pd
from scipy import ndimage
from six.moves.urllib.request import urlretrieve
from six.moves import cPickle as pickle
import tensorflow as tf
print 'tf.__version__:%s' % str(tf.__version__)
%run img_utils.py
Analytics Specs¶
This Project¶
In [3]:
%run img_glbSpec_SFDD_ImgSz_64.py
In [4]:
# print '\nglbDataFile: %s' % (glbDataFile)
print '\nglbRspClass: %s' % (glbRspClass)
print 'glbRspClassN: %d' % (glbRspClassN)
print 'glbRspClassDesc: '; print(glbRspClassDesc)
print '\nglbImg:'; print(glbImg)
print '\nglbTfwVarSeed: %d' % (glbTfwVarSeed)
print '\nglbPickleFile: %s' % (glbPickleFile)
Import Data¶
This Project¶
In [5]:
%run img_utils.py
glbObsFitIdn, glbObsFitFtr, glbObsFitRsp, \
glbObsVldIdn, glbObsVldFtr, glbObsVldRsp, \
glbObsNewIdn, glbObsNewFtr, glbObsNewRsp, \
sbtNewCorDf, \
_ = myimportDbs(glbPickleFile['data'])
glbObsTrnIdn = glbObsFitIdn + glbObsVldIdn
glbObsTrnFtr = np.vstack((glbObsFitFtr, glbObsVldFtr))
glbObsTrnRsp = np.concatenate((glbObsFitRsp, glbObsVldRsp))
print('Fit pickled set:',
len(glbObsFitIdn), glbObsFitFtr.shape, glbObsFitRsp.shape)
print('Vld pickled set:',
len(glbObsVldIdn), glbObsVldFtr.shape, glbObsVldRsp.shape)
print('Trn pickled set:',
len(glbObsTrnIdn), glbObsTrnFtr.shape, glbObsTrnRsp.shape)
print('New pickled set:',
len(glbObsNewIdn), glbObsNewFtr.shape, glbObsNewRsp.shape)
First reload the data we generated in 1_notmist.ipynb.
In [7]:
# pickle_file = 'data/notMNIST.pickle'
# with open(pickle_file, 'rb') as f:
# save = pickle.load(f)
# glbXFit = save['glbXFit']
# glbYFit = save['glbYFit']
# glbXVld = save['glbXVld']
# glbYVld = save['glbYVld']
# glbXNew = save['glbXNew']
# glbYNew = save['glbYNew']
# del save # hint to help gc free up memory
# print('Training set', glbXFit.shape, glbYFit.shape)
# print('Validation set', glbXVld.shape, glbYVld.shape)
# print('Test set', glbXNew.shape, glbYNew.shape)
Reformat into a shape that's more adapted to the models we're going to train:
- data as a flat matrix,
- labels as float 1-hot encodings.
In [6]:
def lclreformatData(I, X, Y):
X = X.reshape((-1, glbImg['size'] * glbImg['size'])).astype(np.float32)
# Map 0 to [1.0, 0.0, 0.0 ...], 1 to [0.0, 1.0, 0.0 ...]
Y = (np.arange(glbRspClassN) == Y[:,None]).astype(np.float32)
return I, X, Y
glbITrn, glbXTrn, glbYTrn = lclreformatData(
glbObsTrnIdn, glbObsTrnFtr, glbObsTrnRsp)
glbIFit, glbXFit, glbYFit = lclreformatData(
glbObsFitIdn, glbObsFitFtr, glbObsFitRsp)
glbIVld, glbXVld, glbYVld = lclreformatData(
glbObsVldIdn, glbObsVldFtr, glbObsVldRsp)
glbINew, glbXNew, glbYNew = lclreformatData(
glbObsNewIdn, glbObsNewFtr, glbObsNewRsp)
print('Trn reshaped set:', len(glbITrn), glbXTrn.shape, glbYTrn.shape)
print('Fit reshaped set:', len(glbIFit), glbXFit.shape, glbYFit.shape)
print('Vld reshaped set:', len(glbIVld), glbXVld.shape, glbYVld.shape)
print('New reshaped set:', len(glbINew), glbXNew.shape, glbYNew.shape)
In [7]:
# Check how much incremental memory is used for Fit obs
del glbObsFitIdn, glbObsFitFtr, glbObsFitRsp
del glbIFit, glbXFit, glbYFit
# Check how much incremental memory is used for Trn obs
# del glbObsTrnIdn, glbObsTrnFtr, glbObsTrnRsp
# del glbITrn, glbXTrn, glbYTrn
In [8]:
print glbObsFitFtr.shape
# print glbObsTrnFtr.shape
In [9]:
# glbImg['size'] = 28
# glbRspClassN = 10
# def reformat(dataset, labels):
# dataset = dataset.reshape((-1, glbImg['size'] * glbImg['size'])).astype(np.float32)
# # Map 0 to [1.0, 0.0, 0.0 ...], 1 to [0.0, 1.0, 0.0 ...]
# labels = (np.arange(glbRspClassN) == labels[:,None]).astype(np.float32)
# return dataset, labels
# glbXFit, glbYFit = reformat(glbXFit, glbYFit)
# glbXVld, glbYVld = reformat(glbXVld, glbYVld)
# glbXNew, glbYNew = reformat(glbXNew, glbYNew)
# print('Training set', glbXFit.shape, glbYFit.shape)
# print('Validation set', glbXVld.shape, glbYVld.shape)
# print('Test set', glbXNew.shape, glbYNew.shape)
Fit Logistic Regression (TensorFlow)¶
We're first going to train a multinomial logistic regression using simple gradient descent.
TensorFlow works like this:
First you describe the computation that you want to see performed: what the inputs, the variables, and the operations look like. These get created as nodes over a computation graph. This description is all contained within the block below:
with graph.as_default(): ...Then you can run the operations on this graph as many times as you want by calling
session.run(), providing it outputs to fetch from the graph that get returned. This runtime operation is all contained in the block below:with tf.Session(graph=graph) as session: ...
Let's load all the data into TensorFlow and build the computation graph corresponding to our training:
In [10]:
print "glbYNew[:5, :]:" % ()
print glbYNew[:5, :]
In [9]:
%run img_glbSpec_SFDD_ImgSz_64.py
%run img_utils.py
# Add parameter for lrnRateTfw = 0.5 for GradientDescentOptimizer
# With gradient descent training, even this much data is prohibitive.
# Subset the training data for faster turnaround.
def fitMdlLgtRgrTfw(lclXFit, lclYFit,
nObsFit = 50, nStepsTfw = 11, lrnRateTfw = 0.5,
visualize = False, newObs = False, verbose = False):
from sklearn import metrics as skl_metrics
print('\nLogistic Regression (TensorFlow): ' + \
"nObsFit:%5d; nStepsTfw:%5d; lrnRateTfw:%.4f" % ( \
nObsFit, nStepsTfw, lrnRateTfw))
print(" visualize: %s; newObs: %s; verbose: %s" % ( \
visualize, newObs, verbose))
startTm = datetime.now()
mdlDf = pd.DataFrame({'id': 'LgtRgr.tfw',
'nObsFit' : [nObsFit],
'nStepsTfw': [nStepsTfw],
'lrnRateTfw': [lrnRateTfw]
})
graph = tf.Graph()
with graph.as_default():
# Input data.
# Load the training, validation and test data into constants that
# are attached to the graph.
tfwXFit = tf.constant(lclXFit[:nObsFit, :])
tfwYFit = tf.constant(lclYFit[:nObsFit])
tfwXVld = tf.constant(glbXVld)
tfwYVld = tf.constant(glbYVld)
# Create placeholders for ObsNew & occlusion visualization
tfwXNew = tf.placeholder(tf.float32,
shape = (glbImg['size'], lclXFit.shape[1]))
tfwYNew = tf.placeholder(tf.float32,
shape = (glbImg['size'], lclYFit.shape[1]))
# Variables.
tf.set_random_seed(glbTfwVarSeed)
# These are the parameters that we are going to be training.
# The weight matrix will be initialized using random valued
# following a (truncated) normal distribution.
# The bias vector get initialized to zero.
tfwW = tf.Variable(
tf.truncated_normal([glbImg['size'] * glbImg['size'],
glbRspClassN]),
name = 'tfwW')
tfwB = tf.Variable(tf.zeros([glbRspClassN]), name = 'tfwB')
if (verbose):
print(' tfwW:', tfwW.initialized_value())
print(' tfwB:', tfwB.initialized_value())
# print 'lblIx:%2d:%s'% \
# (np.vectorize("%.4e".__mod__)(tfwW.value()[:5, lblIx]))
# Training computation.
# We multiply the inputs with the weight matrix, and add bias.
# We compute the softmax and cross-entropy (it's one operation in
# TensorFlow, because it's very common, and it can be optimized).
# We take the average of this cross-entropy across all training
# examples: that's our loss.
logits = tf.matmul(tfwXFit, tfwW) + tfwB
loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(logits, tfwYFit))
# Optimizer.
# We are going to find the minimum of this loss using
# gradient descent.
optimizer = (tf.train
.GradientDescentOptimizer(tf.to_float(lrnRateTfw))
.minimize(loss))
# Predictions for the training, validation, and test data.
# These are not part of training, but merely here so that we can
# report accuracy_score figures as we train.
tfwYFitPby = tf.nn.softmax(logits)
tfwYVldPby = tf.nn.softmax(tf.matmul(tfwXVld, tfwW) + tfwB)
tfwYNewPby = tf.nn.softmax(tf.matmul(tfwXNew, tfwW) + tfwB)
# print ' fitMdlLgtRgrTfw: type(tfwYOccPby): %s' % \
# (type(tfwYOccPby))
# print ' fitMdlLgtRgrTfw: ' + \
# tfwYOccPby.get_shape().as_list(): %s' % \
# (tfwYOccPby.get_shape().as_list())
def accuracy_score(predictions, labels):
return (1.0 * np.sum(np.argmax(predictions, 1) ==
np.argmax(labels, 1))
/ predictions.shape[0])
tf.set_random_seed(glbTfwVarSeed)
with tf.Session(graph=graph) as session:
# This is a one-time operation which ensures the parameters get
# initialized as we described in the graph:
# random tfwW for the matrix, zeros for the tfwB.
tf.initialize_all_variables().run()
if verbose:
print(' Initialized')
# Create a dummy feed for occlusion visualization
# print " fitMdlLgtRgrTfw: lclXFit.shape: %s" % \
# str(lclXFit.shape)
# print " fitMdlLgtRgrTfw: " + \
# "lclXFit[:glbImg['size'], :].shape: %s" % \
# str(lclXFit[:glbImg['size'], :].shape)
feed_dict = {tfwXNew: glbXNew[:glbImg['size'], :]}
for step in range(int(nStepsTfw)):
# Run the computations.
# We tell .run() that we want to run the optimizer,
# and get the loss value and the training predictions returned
# as numpy arrays.
_, l, predictions = \
session.run([optimizer, loss, tfwYFitPby],
feed_dict = feed_dict)
if mydspVerboseTrigger(step):
thsDrn = (datetime.now() - startTm).seconds
if (thsDrn > 100):
print(' logLoss at step %5d: %.4f (%d secs)' % \
(step, l, thsDrn))
accFit = accuracy_score(tfwYFitPby.eval()[:nObsFit, :],
tfwYFit.eval() [:nObsFit, :])
logLossFit = \
skl_metrics.log_loss(tfwYFit.eval() [:nObsFit, :],
tfwYFitPby.eval()[:nObsFit, :])
if verbose:
print '\n Fit accuracy:%0.4f' % (accFit)
print ' Fit logLoss:%0.4f' % (logLossFit)
# print metrics.confusion_matrix(glbYFit[0:nObsFit],
# lclYFitPdn)
# Calling .eval() on tfwObsVldPred is basically like calling run(),
# but just to get that one numpy array.
# Note that it recomputes all its graph dependencies.
lclYVldPby = tfwYVldPby.eval()
lclYVldPdn = np.argmax(lclYVldPby, 1)
mdlDf['accVld'] = accVld = accuracy_score(lclYVldPby, glbYVld)
cnfYVld = skl_metrics.confusion_matrix(glbObsVldRsp, lclYVldPdn)
accYVldCls = cnfYVld.diagonal() * 1.0 / cnfYVld.sum(axis = 1)
mdlDf['accVldCls'] = None
mdlDf.set_value(0, 'accVldCls', {'accCls' : accYVldCls})
mdlDf['logLossVld'] = logLossVld = skl_metrics.log_loss(
glbYVld, lclYVldPby)
logLossVldCls = mygetMetricLogLoss(glbYVld, lclYVldPby,
returnTyp = 'class')
mdlDf['logLossVldCls'] = None
mdlDf.set_value(0, 'logLossVldCls',
{'logLossCls' : logLossVldCls})
if verbose:
print '\n Vld accuracy:%0.4f' % (accVld)
print accYVldCls
print cnfYVld
yLbl = [glbRspClassDesc[glbRspClass[ix]] + ':' + \
glbRspClass[ix] + ':actl' \
for ix in xrange(glbRspClassN)]
xLbl = ['pred:' + glbRspClass[ix] + ':' + \
glbRspClassDesc[glbRspClass[ix]] \
for ix in xrange(glbRspClassN)]
# print labels
plt.matshow(cnfYVld, cmap='Reds', interpolation='none')
plt.yticks(np.arange(10), yLbl)
plt.xticks(np.arange(10), xLbl, rotation=90);
plt.show()
print '\n Vld logLoss:%0.4f' % (logLossVld)
print logLossVldCls
if visualize:
mydisplayImagePredictions(session, tfwW.eval(),
glbIVld, glbObsVldFtr, glbObsVldRsp, lclYVldPby,
glbRspClass, glbRspClassDesc, imgVisualFn = None,
tfwXOcc = tfwXNew, tfwYOccPby = tfwYNewPby)
if newObs:
print " predicting %5d new obs..." % (glbYNew.shape[0])
lclYNewPby = np.zeros((glbYNew.shape[0],
tfwYFitPby.get_shape().as_list()[1]))
lclYNewPby[:, :] = -1.0
btcSz = tfwXNew.get_shape().as_list()[0]
for obsIx in xrange(0, glbYNew.shape[0], btcSz):
if mydspVerboseTrigger(obsIx) and \
(datetime.now() - startTm).seconds > 60:
print " @%5d secs: obsIx: %5d" % \
((datetime.now() - startTm).seconds, obsIx)
obsEnd = obsIx + btcSz
if obsEnd > lclYNewPby.shape[0]:
obsEnd = lclYNewPby.shape[0]
btcYNewPby = session.run(tfwYNewPby,
feed_dict = {tfwXNew: glbXNew[obsIx:obsEnd, :] \
if obsEnd != lclYNewPby.shape[0] \
else np.vstack((glbXNew[obsIx:obsEnd, :],
glbXNew[0:((obsIx + btcSz) % obsEnd), :]))
})
lclYNewPby[obsIx:obsEnd, :] = btcYNewPby[:, :] \
if obsEnd != lclYNewPby.shape[0] \
else btcYNewPby[:(obsEnd - obsIx), :]
assert (lclYNewPby[:, :] != -1.0).all(), \
'some cells in lclYNewPby == -1.0'
# lclYNewPdn = tfwYNew.eval()
# lclYNewPby = tfwYNewPby.eval()
lclYNewPdn = np.argmax(lclYNewPby, 1)
#if (tfwYNew.eval() > -1).any():
if (len(np.unique(glbYNew, return_counts = True)[0]) > 1):
mdlDf['accNew'] = accNew = accuracy_score(lclYNewPby,
glbYNew)
mdlDf['logLossNew'] = logLossNew = skl_metrics.log_loss(
glbYNew, lclYNewPby)
if verbose:
print '\n New accuracy:%0.4f' % (accNew)
print ' New logLoss:%0.4f' % (logLossNew)
print skl_metrics.confusion_matrix(glbObsNewRsp,
lclYNewPdn)
yLbl = [glbRspClassDesc[glbRspClass[ix]] + ':' +
glbRspClass[ix] + ':actl' \
for ix in xrange(glbRspClassN)]
xLbl = ['pred:' + glbRspClass[ix] + ':' + \
glbRspClassDesc[glbRspClass[ix]] \
for ix in xrange(glbRspClassN)]
# print labels
plt.matshow(skl_metrics.confusion_matrix(glbObsNewRsp,
lclYNewPdn),
cmap='Reds', interpolation='none')
plt.yticks(np.arange(10), yLbl)
plt.xticks(np.arange(10), xLbl, rotation=90);
plt.show()
if visualize:
mydisplayImagePredictions(session, tfwW.eval(),
glbINew, glbObsNewFtr, glbObsNewRsp, lclYNewPby,
glbRspClass, glbRspClassDesc, imgVisualFn = None,
tfwXOcc = tfwXNew, tfwYOccPby = tfwYNewPby)
mdlDf['predNew'] = None
mdlDf.set_value(0, 'predNew', {'kntCls' : np.unique(lclYNewPdn,
return_counts = True)})
if verbose:
print '\n New prediction knts:'
print mdlDf['predNew'][0]
# indentation (6 spaces) determines scope of this
# before session.__exit__ & graph.__exit__
mdlDf['model'] = session
mdlDf['duration'] = (datetime.now() - startTm).seconds
print(' duration: %.2d seconds' % (mdlDf['duration'][0]))
if not newObs: lclYNewPby = None
return(mdlDf, lclYVldPby, lclYNewPby)
tmpMdlDf = pd.DataFrame()
# thsMdlDf, thsYVldPby, thsYNewPby = fitMdlLgtRgrTfw(
# glbXFit, glbYFit,
# nObsFit = 100, nStepsTfw = 10, lrnRateTfw = 0.5,
# visualize = True, newObs = True, verbose = True)
# mdlDf = mdlDf.append(thsMdlDf)
# To check if model results are deterministic &
# all run options work separately
# thsMdlDf, thsYVldPby, thsYNewPby = fitMdlLgtRgrTfw(
# glbXFit, glbYFit,
# nObsFit = 100, nStepsTfw = 10, lrnRateTfw = 0.5,
# visualize = True, newObs = False, verbose = False)
# mdlDf = mdlDf.append(thsMdlDf)
thsMdlDf, thsYVldPby, thsYNewPby = fitMdlLgtRgrTfw(
glbXFit, glbYFit,
nObsFit = 100, nStepsTfw = 10, lrnRateTfw = 0.5,
visualize = False, newObs = True, verbose = False)
tmpMdlDf = tmpMdlDf.append(thsMdlDf)
thsMdlDf, thsYVldPby, thsYNewPby = fitMdlLgtRgrTfw(
glbXFit, glbYFit,
nObsFit = 100, nStepsTfw = 10, lrnRateTfw = 0.5,
visualize = False, newObs = False, verbose = True)
tmpMdlDf = tmpMdlDf.append(thsMdlDf)
thsMdlDf, thsYVldPby, thsYNewPby = fitMdlLgtRgrTfw(
glbXFit, glbYFit,
nObsFit = 100, nStepsTfw = 10, lrnRateTfw = 0.5,
visualize = False, newObs = False, verbose = False)
tmpMdlDf = tmpMdlDf.append(thsMdlDf)
# thsMdlDf, thsYVldPby, thsYNewPby = fitMdlLgtRgrTfw(
# glbXFit, glbYFit,
# nObsFit = 100, nStepsTfw = 10, lrnRateTfw = 0.1,
# visualize = False, newObs = False, verbose = True)
# tmpMdlDf = tmpMdlDf.append(thsMdlDf)
print '\ntmpMdlDf: '
print(tmpMdlDf)
In [30]:
lr1MdlDf, lr1YVldPby, lr1YNewPby = fitMdlLgtRgrTfw(
glbXFit, glbYFit,
nObsFit = 1000, nStepsTfw = 100, lrnRateTfw = 0.1,
visualize = False, newObs = True, verbose = True)
In [31]:
lr5MdlDf, lr5YVldPby, lr5YNewPby = fitMdlLgtRgrTfw(
glbXFit, glbYFit,
nObsFit = 1000, nStepsTfw = 100, lrnRateTfw = 0.5,
visualize = False, newObs = True, verbose = True)
In [108]:
%run img_utils.py
from sklearn import metrics as skl_metrics
def lclaccuracy_score(predictions, labels):
return (1.0 * np.sum(np.argmax(predictions, 1) ==
np.argmax(labels, 1))
/ predictions.shape[0])
print "\nlrnRateTfw: 0.1; accVld: %0.4f" % \
(lclaccuracy_score(lr1YVldPby, glbYVld))
cnf1YVld = skl_metrics.confusion_matrix(glbObsVldRsp,
np.argmax(lr1YVldPby, 1))
print cnf1YVld
# print cnf1YVld.sum(axis = 1)
print cnf1YVld.diagonal() * 1.0 / cnf1YVld.sum(axis = 1)
# print cnf1YVld.sum(axis = 1).sum()
lls1YVld = skl_metrics.log_loss(glbYVld, lr1YVldPby)
print "\nlrnRateTfw: 0.1; logLossVld: %0.4f" % (lls1YVld)
# lls1YVldCls = np.array([skl_metrics.log_loss(
# glbYVld[glbYVld[:, clsIx] == 1, :],
# lr1YVldPby[glbYVld[:, clsIx] == 1, :]) \
# for clsIx in range(glbRspClassN)]) / glbRspClassN
# print "lrnRateTfw: 0.1; logLossVld.skl classSum: %0.4f" % (lls1YVld.sum())
# print lls1YVld
lls1YVldCls = mygetMetricLogLoss(glbYVld, lr1YVldPby, returnTyp = 'class')
print "lrnRateTfw: 0.1; logLossVldCls :"; print lls1YVldCls
# print "diff: %.4e" % (np.abs(lls1YVld - lls1YVldCls.sum()))
assert np.abs(lls1YVld - lls1YVldCls.sum()) < 1e-04, \
"logLoss from skl: %.4f vs myCls %.4f does not match" % \
(lls1YVld, lls1YVldCls.sum())
print "\nlrnRateTfw: 0.5; accVld: %0.4f" % \
(lclaccuracy_score(lr5YVldPby, glbYVld))
cnf5YVld = skl_metrics.confusion_matrix(glbObsVldRsp,
np.argmax(lr5YVldPby, 1))
print cnf5YVld
print cnf5YVld.diagonal() * 1.0 / cnf5YVld.sum(axis = 1)
lls5YVld = skl_metrics.log_loss(glbYVld, lr5YVldPby)
print "\nlrnRateTfw: 0.5; logLossVld: %0.4f" % (lls5YVld)
lls5YVldCls = mygetMetricLogLoss(glbYVld, lr5YVldPby, returnTyp = 'class')
print "lrnRateTfw: 0.5; logLossVldCls :"; print lls5YVldCls
assert np.abs(lls5YVld - lls5YVldCls.sum()) < 1e-04, \
"logLoss from skl: %.4f vs myCls %.4f does not match" % \
(lls5YVld, lls5YVldCls.sum())
In [10]:
# glbMdlDf = None
# glbMdlDf = pd.DataFrame()
try:
with open(glbPickleFile['models'], 'rb') as f:
glbMdlDf = pickle.load(f)
assert isinstance(glbMdlDf, pd.DataFrame), \
'type(glbMdlDf): %s, expecting pd.DataFrame' % \
(str(type(glbMdlDf)))
except IOError, e:
print e
print 'file %s not present or not appropriate' % \
(glbPickleFile['models'])
print glbMdlDf
In [12]:
%run img_utils.py
srchParamsDct = {
'nObsFit' : [glbObsTrnFtr.shape[0]],
# 'nObsFit' : [100, 1000, 5000, 10000, glbObsFitFtr.shape[0]]
'nStepsTfw' : [1000],
# 'nStepsTfw' : [1000, 2000],
'lrnRateTfw' : [1.0, 10.0]
# 'lrnRateTfw' : [0.1, 1.0, 10.0]
}
jnk = mysearchParams(fitMdlLgtRgrTfw, srchParamsDct = srchParamsDct,
curResultsDf = glbMdlDf,
mode = 'displayonly',
sort_values = ['nObsFit', 'accVld', 'logLossVld', 'duration'],
sort_ascending = [False , True , False, False],
save_drop_cols = 'model',
save_filepathname = glbPickleFile['models'],
lclXFit = glbXFit, lclYFit = glbYFit)
# thsDf, thsObsVldRspPredProba, thsObsNewRspPredProba = fitMdlLgtRgrTfw(
# glbXFit, glbYFit,
# nObsFit = 100, nStepsTfw = 10, lrnRateTfw = 0.5,
# visualize = False, newObs = False, verbose = False)
In [14]:
%run img_utils.py
glbMdlDf = mysearchParams(fitMdlLgtRgrTfw, srchParamsDct = srchParamsDct,
curResultsDf = glbMdlDf,
mode = 'run',
sort_values = ['nObsFit', 'accVld', 'logLossVld', 'duration'],
sort_ascending = [False , False , True, True],
save_filepathname = glbPickleFile['models'],
save_drop_cols = 'model',
lclXFit = glbXFit, lclYFit = glbYFit)
In [16]:
lrn10Srs = glbMdlDf.ix[('LgtRgr.tfw', 1000.0, 18077.0, 10.0)]
# print lrn10Srs
lrn01Srs = glbMdlDf.ix[('LgtRgr.tfw', 1000.0, 18077.0, 1.0)]
# print lrn01Srs
print "lrn10Srs: accVld: %.4f" % (lrn10Srs['accVld'])
print "lrn01Srs: accVld: %.4f" % (lrn01Srs['accVld'])
print "lrn10-01: accVldDff: %.4f" % (lrn10Srs['accVld'] - lrn01Srs['accVld'])
print "lrn10Srs: accVldCls:"; print (lrn10Srs['accVldCls']['accCls'])
print "lrn01Srs: accVldCls:"; print (lrn01Srs['accVldCls']['accCls'])
print "lrn10-01: accVldClsDff:"; print (lrn10Srs['accVldCls']['accCls'] -
lrn01Srs['accVldCls']['accCls'])
print "\n"
print "lrn10Srs: logLossVld: %8.4f" % (lrn10Srs['logLossVld'])
print "lrn01Srs: logLossVld: %8.4f" % (lrn01Srs['logLossVld'])
print "lrn10-01: logLossVldDff: %8.4f" % (lrn10Srs['logLossVld'] -
lrn01Srs['logLossVld'])
print "lrn10Srs: logLossVldCls:"; print (lrn10Srs['logLossVldCls']['logLossCls'])
print "lrn01Srs: logLossVldCls:"; print (lrn01Srs['logLossVldCls']['logLossCls'])
print "lrn10-01: logLossVldClsDff:";
print (lrn10Srs['logLossVldCls']['logLossCls'] -
lrn01Srs['logLossVldCls']['logLossCls'])
In [158]:
# Set value based on condition
# print glbMdlDf.ix[glbMdlDf['id'].str.contains('LogisticRegression.SGD.tf',
# na=False), 'id']
# glbMdlDf.ix[glbMdlDf['id'].str.contains('LogisticRegression.SGD.tf',
# na=False), 'id'] = 'LgtRgr.SGD.tf'
# print glbMdlDf.ix[glbMdlDf['id'].str.contains('LogisticRegression.SGD.tf',
# na=False), 'id']
def lclfixNanDf(df, column, default):
print df[column].isnull()
df.ix[df[column].isnull(), column] = default
return(df)
# col = 'lrnRateTfw'; default = -1.0;
# tmpMdlDf = lclfixNanDf(glbMdlDf, col, default); print tmpMdlDf[col]
# print glbMdlDf.ix[- glbMdlDf['nStepsTfw'].isnull(), ['id', 'nStepsTfw']]
# glbMdlDf.ix[- glbMdlDf['nStepsTfw'].isnull(), 'id'] = 'LogisticRegression.tf'
# print glbMdlDf.ix[- glbMdlDf['nStepsTfw'].isnull(), ['id', 'nStepsTfw']]
# print glbMdlDf.ix[glbMdlDf['nObsBtc'].isnull(), ['nObsBtc', 'nObsFit']]
# # glbMdlDf.ix[glbMdlDf['nObsBtc'].isnull(), 'nObsBtc'] = \
# # glbMdlDf.ix[glbMdlDf['nObsBtc'].isnull(), 'nObsFit']
# glbMdlDf['nObsBtc'] = glbMdlDf.apply(
# lambda (row): row['nObsFit'] if pd.isnull(row['nObsBtc']) else row['nObsBtc'],
# axis = 1)
# print 'After:'
# print glbMdlDf.ix[glbMdlDf['nObsBtc'].isnull(), ['nObsBtc', 'nObsFit']]
# Change value
# tmpMdlDf = glbMdlDf
# print tmpMdlDf[(tmpMdlDf['id'].str.contains('LgtRgr.skl', na = False)) &
# (tmpMdlDf['nStepsTfw'] == 1.0)]
# print tmpMdlDf.ix[(tmpMdlDf['id'].str.contains('LgtRgr.skl', na = False)) &
# (tmpMdlDf['nStepsTfw'] == 1.0), 'nStepsTfw']
# tmpMdlDf.ix[(tmpMdlDf['id'].str.contains('LgtRgr.skl', na = False)) &
# (tmpMdlDf['nStepsTfw'] == 1.0), 'nStepsTfw'] = -1.0
# print 'After:'
# print tmpMdlDf
# Remove specific models
mask = (glbMdlDf['id'].str.contains('LgtRgr.tfw', na = False))
# mask = (glbMdlDf['id'].str.contains('LgtRgr.tfw', na = False)) & \
# (glbMdlDf['nObsFit'] == 10000.0)
print mask
tmpMdlDf = glbMdlDf[~mask]
# print tmpMdlDf
# Remove dups
# print glbMdlDf.columns
# print (glbMdlDf['logLossVld'])
# print (glbMdlDf.index.duplicated())
# tmpMdlDf = glbMdlDf[~glbMdlDf.index.duplicated()]
# print (tmpMdlDf.index.duplicated())
glbMdlDf = tmpMdlDf
print 'After:'
print glbMdlDf
In [26]:
myexportDf(glbMdlDf,
save_filepathname = glbPickleFile['models'],
save_drop_cols = 'model'
)
In [14]:
glbMdlDf['bestFit'] = False
glbMdlDf.ix[('LgtRgr.tfw', 1000.0, 18077.0, 10.0), 'bestFit'] = True
# LgtRgr.tfw 1000.0 18077.0 10.0
print glbMdlDf[list(set(glbMdlDf.columns) - set(srchParamsDct.keys()))]
# print glbMdlDf[glbMdlDf.nObsFit >= 10000][
# list(set(glbMdlDf.columns) - set(srchParamsDct.keys()))]
In [4]:
# # With gradient descent training, even this much data is prohibitive.
# # Subset the training data for faster turnaround.
# nObsFit = 10000
# graph = tf.Graph()
# with graph.as_default():
# # Input data.
# # Load the training, validation and test data into constants that are
# # attached to the graph.
# tfwObsFitFtr = tf.constant(glbXFit[:nObsFit, :])
# tfwObsFitRsp = tf.constant(glbYFit[:nObsFit])
# tfwObsVldFtr = tf.constant(glbXVld)
# tfwObsNewFtr = tf.constant(glbXNew)
# # Variables.
# # These are the parameters that we are going to be training. The weight
# # matrix will be initialized using random valued following a (truncated)
# # normal distribution. The tfwB get initialized to zero.
# tfwW = tf.Variable(
# tf.truncated_normal([glbImg['size'] * glbImg['size'], glbRspClassN]), name = 'tfwW')
# tfwB = tf.Variable(tf.zeros([glbRspClassN]), name = 'tfwB')
# print(tfwW.initialized_value())
# print(tfwB.initialized_value())
# # print 'initial tfwB:%s' % (np.vectorize("%.4e".__mod__)(tf.get_variable('tfwB')))
# # print 'initial tfwW (first 5 only):'
# # for lblIx in xrange(glbRspClassN):
# # print 'lblIx:%2d:%s'% (np.vectorize("%.4e".__mod__)(tfwW.value()[:5, lblIx]))
# # Training computation.
# # We multiply the inputs with the weight matrix, and add tfwB. We compute
# # the softmax and cross-entropy (it's one operation in TensorFlow, because
# # it's very common, and it can be optimized). We take the average of this
# # cross-entropy across all training examples: that's our loss.
# logits = tf.matmul(tfwObsFitFtr, tfwW) + tfwB
# loss = tf.reduce_mean(
# tf.nn.softmax_cross_entropy_with_logits(logits, tfwObsFitRsp))
# # Optimizer.
# # We are going to find the minimum of this loss using gradient descent.
# optimizer = tf.train.GradientDescentOptimizer(0.5).minimize(loss)
# # Predictions for the training, validation, and test data.
# # These are not part of training, but merely here so that we can report
# # accuracy_score figures as we train.
# tfwObsTrnPred = tf.nn.softmax(logits)
# tfwObsVldPred = tf.nn.softmax(
# tf.matmul(tfwObsVldFtr, tfwW) + tfwB)
# tfwObsNewPred = tf.nn.softmax(tf.matmul(tfwObsNewFtr, tfwW) + tfwB)
Let's run this computation and iterate:
In [58]:
# nStepsTfw = 801
# def accuracy_score(predictions, labels):
# return (100.0 * np.sum(np.argmax(predictions, 1) == np.argmax(labels, 1))
# / predictions.shape[0])
# with tf.Session(graph=graph) as session:
# # This is a one-time operation which ensures the parameters get initialized as
# # we described in the graph: random tfwW for the matrix, zeros for the
# # tfwB.
# tf.initialize_all_variables().run()
# print('Initialized')
# for step in range(nStepsTfw):
# # Run the computations. We tell .run() that we want to run the optimizer,
# # and get the loss value and the training predictions returned as numpy
# # arrays.
# _, l, predictions = session.run([optimizer, loss, tfwObsTrnPred])
# if (step % 100 == 0):
# print('Loss at step %d: %f' % (step, l))
# print('Training accuracy_score: %.1f%%' % accuracy_score(
# predictions, glbYFit[:nObsFit, :]))
# # Calling .eval() on tfwObsVldPred is basically like calling run(), but
# # just to get that one numpy array. Note that it recomputes all its graph
# # dependencies.
# print('Validation accuracy_score: %.1f%%' % accuracy_score(
# tfwObsVldPred.eval(), glbYVld))
# print('Test accuracy_score: %.1f%%' % accuracy_score(tfwObsNewPred.eval(), glbYNew))
In [22]:
robjects.pandas2ri.activate()
pltRDf = robjects.conversion.py2ri(glbMdlDf)
# print(pltRDf)
pltRFn = robjects.r("""
source('~/Dropbox/datascience/R/myplot.R')
function(RDf, filename) {
mypltModelStats(RDf, c('accVld', 'logLossVld', 'duration'),
dim = c('nObsFit', 'id', 'nStepsTfw', 'lrnRateTfw'),
scaleXFn = NULL,
#highLightIx = which.min(RDf$logLossVld),
highLightIx = which(RDf$bestFit == 'TRUE'),
title = NULL,
fileName = filename)
}
""")
pltRFn(pltRDf, 'img_03_fit_lgtRgr_Tfw_SFDD_glbMdlDf.png')
pltRFn = robjects.r("""
source('~/Dropbox/datascience/R/myplot.R')
function(RDf, filename) {
mypltModelStats(RDf, c('accVld', 'logLossVld'),
dim = c('nObsFit', 'id', 'nStepsTfw', 'lrnRateTfw'),
scaleXFn = NULL,
#highLightIx = which.min(RDf$logLossVld),
highLightIx = which(RDf$bestFit == 'TRUE'),
title = NULL,
fileName = filename)
}
""")
pltRFn(pltRDf, 'img_03_fit_lgtRgr_Tfw_SFDD_glbMdlDf_logLossVld.png')
pltRFn = robjects.r("""
source('~/Dropbox/datascience/R/myplot.R')
function(RDf, filename) {
mypltModelStats(RDf, c('accVld'),
dim = c('nObsFit', 'id', 'nStepsTfw', 'lrnRateTfw'),
scaleXFn = NULL,
#highLightIx = which.min(RDf$logLossVld),
highLightIx = which(RDf$bestFit == 'TRUE'),
title = NULL,
fileName = filename)
}
""")
pltRFn(pltRDf, 'img_03_fit_lgtRgr_Tfw_SFDD_glbMdlDf_accVld.png')
Out[22]:
In [21]:
glbMdlDf.to_csv('img_03_fit_lgtRgr_Tfw_SFDD_glbMdlDf.csv')
Fit selected model to glbObsFit¶
In [15]:
selMdlSrs = glbMdlDf[glbMdlDf['bestFit']]
print selMdlSrs
In [17]:
selMdlDf, selYVldPby, selYNewPby = fitMdlLgtRgrTfw(
glbXFit, glbYFit,
nObsFit = glbXFit.shape[0],
nStepsTfw = selMdlSrs['nStepsTfw'][0],
lrnRateTfw = selMdlSrs['lrnRateTfw'][0],
visualize = True, newObs = True, verbose = True)
In [18]:
print '\n selMdl:'
glbwriteSubmission(glbINew, selYNewPby,
'img_03_fit_lgtRgr_Tfw_SFDD_ImgSz_' + str(glbImg['size']) + \
'_sbmt_sel.csv')
Fit selected model to glbObsTrn¶
In [16]:
finMdlDf, finYVldPby, finYNewPby = fitMdlLgtRgrTfw(
glbXTrn, glbYTrn,
nObsFit = glbXTrn.shape[0],
nStepsTfw = selMdlSrs['nStepsTfw'][0],
lrnRateTfw = selMdlSrs['lrnRateTfw'][0],
visualize = True, newObs = True, verbose = True)
# finMdlDf, glbYNewPredProba = fitMdlLgtRgrTfw(glbXTrn, glbYTrn,
# nObsFit = glbXTrn.shape[0],
# nStepsTfw = selMdlSrs['nStepsTfw'][0],
# verbose = True)
In [17]:
print finMdlDf
In [18]:
glbMdlDf = glbMdlDf.append(finMdlDf)
glbMdlDf = glbMdlDf.set_index(['id'] + srchParamsDct.keys(), drop = False)
glbMdlDf = glbMdlDf.sort_values(
['nObsFit', 'accVld', 'logLossVld', 'duration'],
ascending = [False , True , False, False])
print(glbMdlDf[list(set(glbMdlDf.columns) -
set(['id'] + srchParamsDct.keys()))])
In [19]:
myexportDf(glbMdlDf,
save_filepathname = glbPickleFile['models'],
save_drop_cols = 'model'
)
Output submission¶
In [20]:
print '\n finMdl:'
glbwriteSubmission(glbINew, finYNewPby,
'img_03_fit_lgtRgr_Tfw_SFDD_ImgSz_' + str(glbImg['size']) + \
'_sbmt_fin.csv')
In [24]:
prtStr = 'LeaderBoard metric for this sel submission: %0.5f vs. ' + \
'logLossVld (sel): %0.5f'
print prtStr % (20.00136, 19.2099)
prtStr = 'LeaderBoard metric for this fin submission: %0.5f vs. ' + \
'logLossVld (fin): %0.5f'
print prtStr % (18.99503, 0.193149)
print 'Best score yet:%s: %0.5f' % \
('img_02_fit_lgtRgr(Skl)_SFDD_(ImgSz_32_)sbmt(_fin).csv', 2.63892)
Stop here¶
Following code should be in img04_fit_lgtRgrSGDTf
Let's now switch to stochastic gradient descent training instead, which is much faster.
The graph will be similar, except that instead of holding all the training data into a constant node, we create a Placeholder node which will be fed actual data at every call of sesion.run().
In [6]:
import pandas as pd
models = pd.DataFrame({'nRELUs': [0]})
#models.ix[0, 'accuracy_scoreTest'] = 0
print models
In [7]:
batch_size = 128
graph = tf.Graph()
with graph.as_default():
# Input data. For the training data, we use a placeholder that will be fed
# at run time with a training minibatch.
tfwObsFitFtr = tf.placeholder(tf.float32,
shape=(batch_size, glbImg['size'] * glbImg['size']))
tfwObsFitRsp = tf.placeholder(tf.float32, shape=(batch_size, glbRspClassN))
tfwObsVldFtr = tf.constant(glbXVld)
tfwObsNewFtr = tf.constant(glbXNew)
# Variables.
tfwW = tf.Variable(
tf.truncated_normal([glbImg['size'] * glbImg['size'], glbRspClassN]))
tfwB = tf.Variable(tf.zeros([glbRspClassN]))
print(tfwW.initialized_value())
print(tfwB.initialized_value())
# Training computation.
logits = tf.matmul(tfwObsFitFtr, tfwW) + tfwB
loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(logits, tfwObsFitRsp))
# Optimizer.
optimizer = tf.train.GradientDescentOptimizer(0.5).minimize(loss)
# Predictions for the training, validation, and test data.
tfwObsTrnPred = tf.nn.softmax(logits)
tfwObsVldPred = tf.nn.softmax(
tf.matmul(tfwObsVldFtr, tfwW) + tfwB)
tfwObsNewPred = tf.nn.softmax(tf.matmul(tfwObsNewFtr, tfwW) + tfwB)
Let's run it:
In [8]:
nStepsTfw = 3001
with tf.Session(graph=graph) as session:
tf.initialize_all_variables().run()
print("Initialized")
for step in range(nStepsTfw):
# Pick an offset within the training data, which has been randomized.
# Note: we could use better randomization across epochs.
offset = (step * batch_size) % (glbYFit.shape[0] - batch_size)
# Generate a minibatch.
batch_data = glbXFit[offset:(offset + batch_size), :]
batch_labels = glbYFit[offset:(offset + batch_size), :]
# Prepare a dictionary telling the session where to feed the minibatch.
# The key of the dictionary is the placeholder node of the graph to be fed,
# and the value is the numpy array to feed to it.
feed_dict = {tfwObsFitFtr : batch_data, tfwObsFitRsp : batch_labels}
_, l, predictions = session.run(
[optimizer, loss, tfwObsTrnPred], feed_dict=feed_dict)
if (step % 500 == 0):
print("Minibatch loss at step %d: %f" % (step, l))
print("Minibatch accuracy_score: %.1f%%" % accuracy_score(predictions, batch_labels))
print("Validation accuracy_score: %.1f%%" % accuracy_score(
tfwObsVldPred.eval(), glbYVld))
print("Test accuracy_score: %.1f%%" % accuracy_score(tfwObsNewPred.eval(), glbYNew))
models.ix[0, 'accuracy_scoreVld'] = accuracy_score(tfwObsVldPred.eval(), glbYVld)
models.ix[0, 'accuracy_scoreTst'] = accuracy_score( tfwObsNewPred.eval(), glbYNew)
In [9]:
models.ix[0, 'graph'] = graph
print(models)
Problem¶
Turn the logistic regression example with SGD into a 1-hidden layer neural network with rectified linear units (nn.relu()) and 1024 hidden nodes. This model should improve your validation / test accuracy_score.
In [10]:
nRELUs = [2 ** thsRelu for thsRelu in xrange(11)]
print(nRELUs)
for thsRelu in nRELUs:
models.ix[thsRelu, 'nRELUs'] = thsRelu
print models
In [49]:
thsRelu = nRELUs[9]
batch_size = 128
graph = tf.Graph()
with graph.as_default():
# Input data. For the training data, we use a placeholder that will be fed
# at run time with a training minibatch.
tfwObsFitFtr = tf.placeholder(tf.float32,
shape=(batch_size, glbImg['size'] * glbImg['size']))
tfwObsFitRsp = tf.placeholder(tf.float32, shape=(batch_size, glbRspClassN))
tfwObsVldFtr = tf.constant(glbXVld)
tfwObsNewFtr = tf.constant(glbXNew)
# Variables.
tfwW1 = tf.Variable(
tf.truncated_normal([glbImg['size'] * glbImg['size'], thsRelu]), name = 'tfwW1')
tfwB1 = tf.Variable(tf.zeros([thsRelu]), name = 'tfwB1')
tfwW2 = tf.Variable(
tf.truncated_normal([thsRelu, glbRspClassN]), name = 'tfwW2')
tfwB2 = tf.Variable(tf.zeros([glbRspClassN]), name = 'tfwB2')
print(tfwW1.initialized_value())
print(tfwB1.initialized_value())
#print(relus.initialized_value())
print(tfwW2.initialized_value())
print(tfwB2.initialized_value())
#tf.Print(relus, [relus])
# Training computation.
layer1 = tf.matmul(tfwObsFitFtr, tfwW1) + tfwB1
layer2 = tf.nn.relu(layer1)
layer3 = tf.matmul(layer2, tfwW2) + tfwB2
loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(layer3, tfwObsFitRsp))
# Optimizer.
optimizer = tf.train.GradientDescentOptimizer(0.5).minimize(loss)
# Predictions for the training, validation, and test data.
tfwObsTrnPred = tf.nn.softmax(layer3)
tfwObsVldPred = tf.nn.softmax(
tf.matmul(tf.nn.relu(tf.matmul(tfwObsVldFtr, tfwW1) + tfwB1), tfwW2) + tfwB2)
tfwObsNewPred = tf.nn.softmax(
tf.matmul(tf.nn.relu(tf.matmul(tfwObsNewFtr, tfwW1) + tfwB1), tfwW2) + tfwB2)
In [50]:
nStepsTfw = 3001
with tf.Session(graph=graph) as session:
tf.initialize_all_variables().run()
print("Initialized")
for step in range(nStepsTfw):
# Pick an offset within the training data, which has been randomized.
# Note: we could use better randomization across epochs.
offset = (step * batch_size) % (glbYFit.shape[0] - batch_size)
# Generate a minibatch.
batch_data = glbXFit[offset:(offset + batch_size), :]
batch_labels = glbYFit[offset:(offset + batch_size), :]
# Prepare a dictionary telling the session where to feed the minibatch.
# The key of the dictionary is the placeholder node of the graph to be fed,
# and the value is the numpy array to feed to it.
feed_dict = {tfwObsFitFtr : batch_data, tfwObsFitRsp : batch_labels}
_, l, predictions = session.run(
[optimizer, loss, tfwObsTrnPred], feed_dict=feed_dict)
if (step % 500 == 0):
print("Minibatch loss at step %d: %f" % (step, l))
print("Minibatch accuracy_score: %.1f%%" % accuracy_score(predictions, batch_labels))
print("Validation accuracy_score: %.1f%%" % accuracy_score(
tfwObsVldPred.eval(), glbYVld))
print("Test accuracy_score: %.1f%%" % accuracy_score(tfwObsNewPred.eval(), glbYNew))
models.ix[thsRelu, 'accuracy_scoreVld'] = accuracy_score(tfwObsVldPred.eval(), glbYVld)
models.ix[thsRelu, 'accuracy_scoreTst'] = accuracy_score( tfwObsNewPred.eval(), glbYNew)
models.ix[thsRelu, 'graph'] = graph
print(models)
In [52]:
plt.figure()
#plt.plot(models['nRELUs'], models['accuracy_score.fit'], 'bo-', label = 'fit')
plt.plot(models['nRELUs'], models['accuracy_scoreVld'], 'rs-', label = 'vld')
plt.plot(models['nRELUs'], models['accuracy_scoreTst'], 'gp-', label = 'new')
plt.legend(loc = 'lower right')
plt.title("accuracy_score")
plt.xscale('symlog', basex=2)
axes = plt.gca()
axes.set_xlabel('nRELUs')
# axes.set_xlim([mdlDF['l1_penalty'][mdlDF['RSS.vld'].argmin()] / 10 ** 2, \
# mdlDF['l1_penalty'][mdlDF['RSS.vld'].argmin()] * 10 ** 2])
# axes.set_ylim([0, mdlDF['RSS.vld'].min() * 1.5])
plt.show()
In [ ]: